aboutsummaryrefslogtreecommitdiffstats
path: root/litmus/nvidia_info.c
diff options
context:
space:
mode:
Diffstat (limited to 'litmus/nvidia_info.c')
-rw-r--r--litmus/nvidia_info.c1137
1 files changed, 1137 insertions, 0 deletions
diff --git a/litmus/nvidia_info.c b/litmus/nvidia_info.c
new file mode 100644
index 000000000000..5a63fb732e8b
--- /dev/null
+++ b/litmus/nvidia_info.c
@@ -0,0 +1,1137 @@
1#include <linux/module.h>
2#include <linux/semaphore.h>
3#include <linux/pci.h>
4
5#include <litmus/sched_trace.h>
6#include <litmus/nvidia_info.h>
7#include <litmus/litmus.h>
8
9#include <litmus/sched_plugin.h>
10
11#include <litmus/binheap.h>
12
13#ifdef CONFIG_LITMUS_SOFTIRQD
14#include <litmus/litmus_softirq.h>
15#endif
16
17typedef unsigned char NvV8; /* "void": enumerated or multiple fields */
18typedef unsigned short NvV16; /* "void": enumerated or multiple fields */
19typedef unsigned char NvU8; /* 0 to 255 */
20typedef unsigned short NvU16; /* 0 to 65535 */
21typedef signed char NvS8; /* -128 to 127 */
22typedef signed short NvS16; /* -32768 to 32767 */
23typedef float NvF32; /* IEEE Single Precision (S1E8M23) */
24typedef double NvF64; /* IEEE Double Precision (S1E11M52) */
25typedef unsigned int NvV32; /* "void": enumerated or multiple fields */
26typedef unsigned int NvU32; /* 0 to 4294967295 */
27typedef unsigned long long NvU64; /* 0 to 18446744073709551615 */
28typedef union
29{
30 volatile NvV8 Reg008[1];
31 volatile NvV16 Reg016[1];
32 volatile NvV32 Reg032[1];
33} litmus_nv_hwreg_t, * litmus_nv_phwreg_t;
34
35typedef struct
36{
37 NvU64 address;
38#ifdef CONFIG_CUDA_5_0
39 NvU64 strapped_size;
40#endif
41 NvU64 size;
42 NvU32 offset;
43 NvU32 *map;
44 litmus_nv_phwreg_t map_u;
45} litmus_nv_aperture_t;
46
47typedef struct
48{
49 void *priv; /* private data */
50 void *os_state; /* os-specific device state */
51
52#ifndef CONFIG_CUDA_5_0
53 int rmInitialized;
54#endif
55 int flags;
56
57 /* PCI config info */
58 NvU32 domain;
59 NvU16 bus;
60 NvU16 slot;
61 NvU16 vendor_id;
62 NvU16 device_id;
63 NvU16 subsystem_id;
64 NvU32 gpu_id;
65 void *handle;
66
67 NvU32 pci_cfg_space[16];
68
69 /* physical characteristics */
70 litmus_nv_aperture_t bars[3];
71 litmus_nv_aperture_t *regs;
72 litmus_nv_aperture_t *fb, ud;
73 litmus_nv_aperture_t agp;
74
75 NvU32 interrupt_line;
76
77 NvU32 agp_config;
78 NvU32 agp_status;
79
80 NvU32 primary_vga;
81
82 NvU32 sim_env;
83
84 NvU32 rc_timer_enabled;
85
86 /* list of events allocated for this device */
87 void *event_list;
88
89 void *kern_mappings;
90
91} litmus_nv_state_t;
92
93typedef struct work_struct litmus_nv_task_t;
94
95typedef struct litmus_nv_work_s {
96 litmus_nv_task_t task;
97 void *data;
98} litmus_nv_work_t;
99
100typedef struct litmus_nv_linux_state_s {
101 litmus_nv_state_t nv_state;
102 atomic_t usage_count;
103
104 struct pci_dev *dev;
105 void *agp_bridge;
106 void *alloc_queue;
107
108 void *timer_sp;
109 void *isr_sp;
110 void *pci_cfgchk_sp;
111 void *isr_bh_sp;
112
113#if defined(CONFIG_CUDA_4_0) || defined(CONFIG_CUDA_5_0)
114 char registry_keys[512];
115#endif
116
117 /* keep track of any pending bottom halfes */
118 struct tasklet_struct tasklet;
119 litmus_nv_work_t work;
120
121 /* get a timer callback every second */
122 struct timer_list rc_timer;
123
124 /* lock for linux-specific data, not used by core rm */
125 struct semaphore ldata_lock;
126
127 /* lock for linux-specific alloc queue */
128 struct semaphore at_lock;
129
130#if 0
131#if defined(NV_USER_MAP)
132 /* list of user mappings */
133 struct nv_usermap_s *usermap_list;
134
135 /* lock for VMware-specific mapping list */
136 struct semaphore mt_lock;
137#endif /* defined(NV_USER_MAP) */
138#if defined(NV_PM_SUPPORT_OLD_STYLE_APM)
139 void *apm_nv_dev;
140#endif
141#endif
142
143 NvU32 device_num;
144 struct litmus_nv_linux_state_s *next;
145} litmus_nv_linux_state_t;
146
147void dump_nvidia_info(const struct tasklet_struct *t)
148{
149 litmus_nv_state_t* nvstate = NULL;
150 litmus_nv_linux_state_t* linuxstate = NULL;
151 struct pci_dev* pci = NULL;
152
153 nvstate = (litmus_nv_state_t*)(t->data);
154
155 if(nvstate)
156 {
157 TRACE("NV State:\n"
158 "\ttasklet ptr = %p\n"
159 "\tstate ptr = %p\n"
160 "\tprivate data ptr = %p\n"
161 "\tos state ptr = %p\n"
162 "\tdomain = %u\n"
163 "\tbus = %u\n"
164 "\tslot = %u\n"
165 "\tvender_id = %u\n"
166 "\tdevice_id = %u\n"
167 "\tsubsystem_id = %u\n"
168 "\tgpu_id = %u\n"
169 "\tinterrupt_line = %u\n",
170 t,
171 nvstate,
172 nvstate->priv,
173 nvstate->os_state,
174 nvstate->domain,
175 nvstate->bus,
176 nvstate->slot,
177 nvstate->vendor_id,
178 nvstate->device_id,
179 nvstate->subsystem_id,
180 nvstate->gpu_id,
181 nvstate->interrupt_line);
182
183 linuxstate = container_of(nvstate, litmus_nv_linux_state_t, nv_state);
184 }
185 else
186 {
187 TRACE("INVALID NVSTATE????\n");
188 }
189
190 if(linuxstate)
191 {
192 int ls_offset = (void*)(&(linuxstate->device_num)) - (void*)(linuxstate);
193 int ns_offset_raw = (void*)(&(linuxstate->device_num)) - (void*)(&(linuxstate->nv_state));
194 int ns_offset_desired = (void*)(&(linuxstate->device_num)) - (void*)(nvstate);
195
196
197 TRACE("LINUX NV State:\n"
198 "\tlinux nv state ptr: %p\n"
199 "\taddress of tasklet: %p\n"
200 "\taddress of work: %p\n"
201 "\tusage_count: %d\n"
202 "\tdevice_num: %u\n"
203 "\ttasklet addr == this tasklet: %d\n"
204 "\tpci: %p\n",
205 linuxstate,
206 &(linuxstate->tasklet),
207 &(linuxstate->work),
208 atomic_read(&(linuxstate->usage_count)),
209 linuxstate->device_num,
210 (t == &(linuxstate->tasklet)),
211 linuxstate->dev);
212
213 pci = linuxstate->dev;
214
215 TRACE("Offsets:\n"
216 "\tOffset from LinuxState: %d, %x\n"
217 "\tOffset from NVState: %d, %x\n"
218 "\tOffset from parameter: %d, %x\n"
219 "\tdevice_num: %u\n",
220 ls_offset, ls_offset,
221 ns_offset_raw, ns_offset_raw,
222 ns_offset_desired, ns_offset_desired,
223 *((u32*)((void*)nvstate + ns_offset_desired)));
224 }
225 else
226 {
227 TRACE("INVALID LINUXNVSTATE?????\n");
228 }
229
230#if 0
231 if(pci)
232 {
233 TRACE("PCI DEV Info:\n"
234 "pci device ptr: %p\n"
235 "\tdevfn = %d\n"
236 "\tvendor = %d\n"
237 "\tdevice = %d\n"
238 "\tsubsystem_vendor = %d\n"
239 "\tsubsystem_device = %d\n"
240 "\tslot # = %d\n",
241 pci,
242 pci->devfn,
243 pci->vendor,
244 pci->device,
245 pci->subsystem_vendor,
246 pci->subsystem_device,
247 pci->slot->number);
248 }
249 else
250 {
251 TRACE("INVALID PCIDEV PTR?????\n");
252 }
253#endif
254}
255
256
257
258static struct module* nvidia_mod = NULL;
259
260
261
262
263#if 0
264static int nvidia_ready_module_notify(struct notifier_block *self,
265 unsigned long val, void *data)
266{
267 mutex_lock(&module_mutex);
268 nvidia_mod = find_module("nvidia");
269 mutex_unlock(&module_mutex);
270
271 if(nvidia_mod != NULL)
272 {
273 TRACE("%s : Found NVIDIA module. Core Code: %p to %p\n", __FUNCTION__,
274 (void*)(nvidia_mod->module_core),
275 (void*)(nvidia_mod->module_core) + nvidia_mod->core_size);
276 init_nv_device_reg();
277 return(0);
278 }
279 else
280 {
281 TRACE("%s : Could not find NVIDIA module! Loaded?\n", __FUNCTION__);
282 }
283}
284
285static int nvidia_going_module_notify(struct notifier_block *self,
286 unsigned long val, void *data)
287{
288 nvidia_mod = NULL;
289 mb();
290
291 return 0;
292}
293
294static struct notifier_block nvidia_ready = {
295 .notifier_call = nvidia_ready_module_notify,
296 .priority = 1,
297};
298
299static struct notifier_block nvidia_going = {
300 .notifier_call = nvidia_going_module_notify,
301 .priority = 1,
302};
303#endif
304
305
306
307static int init_nv_device_reg(void);
308static int shutdown_nv_device_reg(void);
309
310
311int init_nvidia_info(void)
312{
313 mutex_lock(&module_mutex);
314 nvidia_mod = find_module("nvidia");
315 mutex_unlock(&module_mutex);
316 if(nvidia_mod != NULL)
317 {
318 TRACE("%s : Found NVIDIA module. Core Code: %p to %p\n", __FUNCTION__,
319 (void*)(nvidia_mod->module_core),
320 (void*)(nvidia_mod->module_core) + nvidia_mod->core_size);
321 init_nv_device_reg();
322 return(0);
323 }
324 else
325 {
326 TRACE("%s : Could not find NVIDIA module! Loaded?\n", __FUNCTION__);
327
328 init_nv_device_reg();
329 return(0);
330// return(-1);
331 }
332}
333
334void shutdown_nvidia_info(void)
335{
336 nvidia_mod = NULL;
337 mb();
338
339 shutdown_nv_device_reg();
340}
341
342/* works with pointers to static data inside the module too. */
343int is_nvidia_func(void* func_addr)
344{
345 int ret = 0;
346 if(nvidia_mod)
347 {
348 ret = within_module_core((long unsigned int)func_addr, nvidia_mod);
349 /*
350 if(ret)
351 {
352 TRACE("%s : %p is in NVIDIA module: %d\n",
353 __FUNCTION__, func_addr, ret);
354 }*/
355 }
356
357 return(ret);
358}
359
360u32 get_tasklet_nv_device_num(const struct tasklet_struct *t)
361{
362 // life is too short to use hard-coded offsets. update this later.
363 litmus_nv_state_t* nvstate = (litmus_nv_state_t*)(t->data);
364 litmus_nv_linux_state_t* linuxstate = container_of(nvstate, litmus_nv_linux_state_t, nv_state);
365
366 BUG_ON(linuxstate->device_num >= NV_DEVICE_NUM);
367
368 return(linuxstate->device_num);
369}
370
371u32 get_work_nv_device_num(const struct work_struct *t)
372{
373 // offset determined though observed behavior of the NV driver.
374 const int DEVICE_NUM_OFFSET = sizeof(struct work_struct);
375 void* state = (void*)(t);
376 void** device_num_ptr = state + DEVICE_NUM_OFFSET;
377 return(*((u32*)(*device_num_ptr)));
378}
379
380
381///////////////////////////////////////////////////////////////////////////////
382///////////////////////////////////////////////////////////////////////////////
383///////////////////////////////////////////////////////////////////////////////
384
385
386typedef struct {
387 raw_spinlock_t lock; /* not needed if GPU not shared between scheudling domains */
388 struct binheap owners;
389
390#ifdef CONFIG_LITMUS_SOFTIRQD
391 klmirqd_callback_t callback;
392 struct task_struct* thread;
393 int ready:1; /* todo: make threads check for the ready flag */
394#endif
395
396#ifdef CONFIG_LITMUS_NV_KLMIRQD_DEBUG
397 struct tasklet_struct nv_klmirqd_dbg_tasklet;
398#endif
399}nv_device_registry_t;
400
401
402static nv_device_registry_t NV_DEVICE_REG[NV_DEVICE_NUM];
403
404
405
406#ifdef CONFIG_LITMUS_SOFTIRQD
407static int nvidia_klmirqd_cb(void *arg)
408{
409 unsigned long flags;
410 int reg_device_id = (int)(long long)(arg);
411 nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id];
412
413 TRACE("nv klmirqd callback for GPU %d\n", reg_device_id);
414
415 raw_spin_lock_irqsave(&reg->lock, flags);
416 reg->thread = current;
417 reg->ready = 1;
418 raw_spin_unlock_irqrestore(&reg->lock, flags);
419
420 return 0;
421}
422#endif
423
424#ifdef CONFIG_LITMUS_NV_KLMIRQD_DEBUG
425struct nv_klmirqd_dbg_timer_struct
426{
427 struct hrtimer timer;
428};
429
430static struct nv_klmirqd_dbg_timer_struct nv_klmirqd_dbg_timer;
431
432static void nv_klmirqd_arm_dbg_timer(lt_t relative_time)
433{
434 lt_t when_to_fire = litmus_clock() + relative_time;
435
436 TRACE("next nv tasklet in %d ns\n", relative_time);
437
438 __hrtimer_start_range_ns(&nv_klmirqd_dbg_timer.timer,
439 ns_to_ktime(when_to_fire),
440 0,
441 HRTIMER_MODE_ABS_PINNED,
442 0);
443}
444
445static void nv_klmirqd_dbg_tasklet_func(unsigned long arg)
446{
447 lt_t now = litmus_clock();
448 nv_device_registry_t *reg = (nv_device_registry_t*)arg;
449 int gpunum = reg - &NV_DEVICE_REG[0];
450
451 TRACE("nv klmirqd routine invoked for GPU %d!\n", gpunum);
452
453 /* set up the next timer */
454 nv_klmirqd_arm_dbg_timer(now % (NSEC_PER_MSEC * 10)); // within the next 10ms.
455}
456
457
458static enum hrtimer_restart nvklmirqd_timer_func(struct hrtimer *timer)
459{
460 lt_t now = litmus_clock();
461 int gpu = (int)(now % num_online_gpus());
462 nv_device_registry_t *reg;
463
464 TRACE("nvklmirqd_timer invoked!\n");
465
466 reg = &NV_DEVICE_REG[gpu];
467
468 if (reg->thread && reg->ready) {
469 TRACE("Adding a tasklet for GPU %d\n", gpu);
470 litmus_tasklet_schedule(&reg->nv_klmirqd_dbg_tasklet, reg->thread);
471 }
472 else {
473 TRACE("nv klmirqd is not ready!\n");
474 nv_klmirqd_arm_dbg_timer(now % (NSEC_PER_MSEC * 10)); // within the next 10ms.
475 }
476
477 return HRTIMER_NORESTART;
478}
479#endif
480
481
482static int gpu_owner_max_priority_order(struct binheap_node *a,
483 struct binheap_node *b)
484{
485 struct task_struct *d_a = container_of(binheap_entry(a, struct rt_param, gpu_owner_node),
486 struct task_struct, rt_param);
487 struct task_struct *d_b = container_of(binheap_entry(b, struct rt_param, gpu_owner_node),
488 struct task_struct, rt_param);
489
490 BUG_ON(!d_a);
491 BUG_ON(!d_b);
492
493 return litmus->compare(d_a, d_b);
494}
495
496static int init_nv_device_reg(void)
497{
498 int i;
499 char name[MAX_KLMIRQD_NAME_LEN+1];
500
501#ifdef CONFIG_LITMUS_SOFTIRQD
502 if (!klmirqd_is_ready()) {
503 TRACE("klmirqd is not ready!\n");
504 return 0;
505 }
506#endif
507
508 memset(NV_DEVICE_REG, 0, sizeof(NV_DEVICE_REG));
509 mb();
510
511
512 for(i = 0; i < num_online_gpus(); ++i) {
513 raw_spin_lock_init(&NV_DEVICE_REG[i].lock);
514 INIT_BINHEAP_HANDLE(&NV_DEVICE_REG[i].owners, gpu_owner_max_priority_order);
515
516#ifdef CONFIG_LITMUS_NV_KLMIRQD_DEBUG
517 tasklet_init(&NV_DEVICE_REG[i].nv_klmirqd_dbg_tasklet, nv_klmirqd_dbg_tasklet_func, (unsigned long)&NV_DEVICE_REG[i]);
518#endif
519
520#ifdef CONFIG_LITMUS_SOFTIRQD
521 {
522 int default_cpu = litmus->map_gpu_to_cpu(i);
523
524 snprintf(name, MAX_KLMIRQD_NAME_LEN, "nvklmirqd%d", i);
525
526 NV_DEVICE_REG[i].callback.func = nvidia_klmirqd_cb;
527 NV_DEVICE_REG[i].callback.arg = (void*)(long long)(i);
528 mb();
529
530 if(launch_klmirqd_thread(name, default_cpu, &NV_DEVICE_REG[i].callback) != 0) {
531 TRACE("Failed to create klmirqd thread for GPU %d\n", i);
532 }
533 }
534#endif
535 }
536
537#ifdef CONFIG_LITMUS_NV_KLMIRQD_DEBUG
538 hrtimer_init(&nv_klmirqd_dbg_timer.timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
539 nv_klmirqd_dbg_timer.timer.function = nvklmirqd_timer_func;
540 nv_klmirqd_arm_dbg_timer(NSEC_PER_MSEC * 1000);
541#endif
542
543 return(1);
544}
545
546
547/* The following code is full of nasty race conditions... */
548/* spawning of klimirqd threads can race with init_nv_device_reg()!!!! */
549static int shutdown_nv_device_reg(void)
550{
551 TRACE("Shutting down nv device registration.\n");
552
553#ifdef CONFIG_LITMUS_SOFTIRQD
554 {
555 int i;
556 nv_device_registry_t *reg;
557
558 for (i = 0; i < num_online_gpus(); ++i) {
559
560 TRACE("Shutting down GPU %d.\n", i);
561
562 reg = &NV_DEVICE_REG[i];
563
564 if (reg->thread && reg->ready) {
565 kill_klmirqd_thread(reg->thread);
566
567 /* assume that all goes according to plan... */
568 reg->thread = NULL;
569 reg->ready = 0;
570 }
571
572 while (!binheap_empty(&reg->owners)) {
573 binheap_delete_root(&reg->owners, struct rt_param, gpu_owner_node);
574 }
575 }
576 }
577#endif
578
579 return(1);
580}
581
582
583/* use to get the owner of nv_device_id. */
584struct task_struct* get_nv_max_device_owner(u32 target_device_id)
585{
586 struct task_struct *owner = NULL;
587 nv_device_registry_t *reg;
588
589 BUG_ON(target_device_id >= NV_DEVICE_NUM);
590
591 reg = &NV_DEVICE_REG[target_device_id];
592
593 if (!binheap_empty(&reg->owners)) {
594 struct task_struct *hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
595 struct task_struct, rt_param);
596 TRACE_CUR("hp: %s/%d\n", hp->comm, hp->pid);
597 }
598
599 return(owner);
600}
601
602#ifdef CONFIG_LITMUS_SOFTIRQD
603struct task_struct* get_nv_klmirqd_thread(u32 target_device_id)
604{
605 struct task_struct *klmirqd = NULL;
606 nv_device_registry_t *reg;
607
608 BUG_ON(target_device_id >= NV_DEVICE_NUM);
609
610 reg = &NV_DEVICE_REG[target_device_id];
611
612 if(likely(reg->ready)) {
613 klmirqd = reg->thread;
614 }
615
616 return klmirqd;
617}
618#endif
619
620
621
622
623
624#ifdef CONFIG_LITMUS_SOFTIRQD
625static int gpu_klmirqd_increase_priority(struct task_struct *klmirqd, struct task_struct *hp)
626{
627 int retval = 0;
628
629 TRACE_CUR("Increasing priority of nv klmirqd: %s/%d.\n", klmirqd->comm, klmirqd->pid);
630
631 /* the klmirqd thread should never attempt to hold a litmus-level real-time
632 * so nested support is not required */
633 retval = litmus->__increase_prio(klmirqd, hp);
634
635 return retval;
636}
637
638static int gpu_klmirqd_decrease_priority(struct task_struct *klmirqd, struct task_struct *hp)
639{
640 int retval = 0;
641
642 TRACE_CUR("Decreasing priority of nv klmirqd: %s/%d.\n", klmirqd->comm, klmirqd->pid);
643
644 /* the klmirqd thread should never attempt to hold a litmus-level real-time
645 * so nested support is not required */
646 retval = litmus->__decrease_prio(klmirqd, hp);
647
648 return retval;
649}
650#endif
651
652
653
654
655/* call when an gpu owner becomes real-time */
656long enable_gpu_owner(struct task_struct *t)
657{
658 long retval = 0;
659// unsigned long flags;
660 int gpu;
661 nv_device_registry_t *reg;
662
663#ifdef CONFIG_LITMUS_SOFTIRQD
664 struct task_struct *hp;
665#endif
666
667 if (!tsk_rt(t)->held_gpus) {
668 TRACE_CUR("task %s/%d does not hold any GPUs\n", t->comm, t->pid);
669 return -1;
670 }
671
672 BUG_ON(!is_realtime(t));
673
674 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
675
676 if (binheap_is_in_heap(&tsk_rt(t)->gpu_owner_node)) {
677 TRACE_CUR("task %s/%d is already active on GPU %d\n", t->comm, t->pid, gpu);
678 goto out;
679 }
680
681 /* update the registration (and maybe klmirqd) */
682 reg = &NV_DEVICE_REG[gpu];
683
684// raw_spin_lock_irqsave(&reg->lock, flags);
685
686 binheap_add(&tsk_rt(t)->gpu_owner_node, &reg->owners,
687 struct rt_param, gpu_owner_node);
688
689
690#ifdef CONFIG_LITMUS_SOFTIRQD
691 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
692 struct task_struct, rt_param);
693
694 if (hp == t) {
695 /* we're the new hp */
696 TRACE_CUR("%s/%d is new hp on GPU %d.\n", t->comm, t->pid, gpu);
697
698 retval = gpu_klmirqd_increase_priority(reg->thread, (tsk_rt(hp)->inh_task)? tsk_rt(hp)->inh_task : hp);
699 }
700#endif
701
702// raw_spin_unlock_irqsave(&reg->lock, flags);
703
704out:
705 return retval;
706}
707
708/* call when an gpu owner exits real-time */
709long disable_gpu_owner(struct task_struct *t)
710{
711 long retval = 0;
712// unsigned long flags;
713 int gpu;
714 nv_device_registry_t *reg;
715
716#ifdef CONFIG_LITMUS_SOFTIRQD
717 struct task_struct *hp;
718 struct task_struct *new_hp = NULL;
719#endif
720
721 if (!tsk_rt(t)->held_gpus) {
722 TRACE_CUR("task %s/%d does not hold any GPUs\n", t->comm, t->pid);
723 return -1;
724 }
725
726 BUG_ON(!is_realtime(t));
727
728 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
729
730 if (!binheap_is_in_heap(&tsk_rt(t)->gpu_owner_node)) {
731 TRACE_CUR("task %s/%d is not active on GPU %d\n", t->comm, t->pid, gpu);
732 goto out;
733 }
734
735 TRACE_CUR("task %s/%d exiting from GPU %d.\n", t->comm, t->pid, gpu);
736
737
738 reg = &NV_DEVICE_REG[gpu];
739
740// raw_spin_lock_irqsave(&reg->lock, flags);
741
742
743#ifdef CONFIG_LITMUS_SOFTIRQD
744 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
745 struct task_struct, rt_param);
746
747 binheap_delete(&tsk_rt(t)->gpu_owner_node, &reg->owners);
748
749
750 if (!binheap_empty(&reg->owners)) {
751 new_hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
752 struct task_struct, rt_param);
753 }
754
755 if (hp == t && new_hp != t) {
756 struct task_struct *to_inh = NULL;
757
758 TRACE_CUR("%s/%d is no longer hp on GPU %d.\n", t->comm, t->pid, gpu);
759
760 if (new_hp) {
761 to_inh = (tsk_rt(new_hp)->inh_task) ? tsk_rt(new_hp)->inh_task : new_hp;
762 }
763
764 retval = gpu_klmirqd_decrease_priority(reg->thread, to_inh);
765 }
766#else
767 binheap_delete(&tsk_rt(t)->gpu_owner_node, &reg->owners);
768#endif
769
770// raw_spin_unlock_irqsave(&reg->lock, flags);
771
772
773out:
774 return retval;
775}
776
777
778
779
780
781
782
783
784
785
786int gpu_owner_increase_priority(struct task_struct *t)
787{
788 int retval = 0;
789 int gpu;
790 nv_device_registry_t *reg;
791
792 struct task_struct *hp = NULL;
793 struct task_struct *hp_eff = NULL;
794
795 BUG_ON(!is_realtime(t));
796 BUG_ON(!tsk_rt(t)->held_gpus);
797
798 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
799
800 if (!binheap_is_in_heap(&tsk_rt(t)->gpu_owner_node)) {
801 WARN_ON(!is_running(t));
802 TRACE_CUR("gpu klmirqd may not inherit from %s/%d on GPU %d\n",
803 t->comm, t->pid, gpu);
804 goto out;
805 }
806
807
808
809
810 TRACE_CUR("task %s/%d on GPU %d increasing priority.\n", t->comm, t->pid, gpu);
811 reg = &NV_DEVICE_REG[gpu];
812
813 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
814 struct task_struct, rt_param);
815 hp_eff = effective_priority(hp);
816
817 if (hp != t) { /* our position in the heap may have changed. hp is already at the root. */
818 binheap_decrease(&tsk_rt(t)->gpu_owner_node, &reg->owners);
819 }
820
821 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
822 struct task_struct, rt_param);
823
824 if (effective_priority(hp) != hp_eff) { /* the eff. prio. of hp has changed */
825 hp_eff = effective_priority(hp);
826 TRACE_CUR("%s/%d is new hp on GPU %d.\n", t->comm, t->pid, gpu);
827
828 retval = gpu_klmirqd_increase_priority(reg->thread, hp_eff);
829 }
830
831out:
832 return retval;
833}
834
835
836int gpu_owner_decrease_priority(struct task_struct *t)
837{
838 int retval = 0;
839 int gpu;
840 nv_device_registry_t *reg;
841
842 struct task_struct *hp = NULL;
843 struct task_struct *hp_eff = NULL;
844
845 BUG_ON(!is_realtime(t));
846 BUG_ON(!tsk_rt(t)->held_gpus);
847
848 gpu = find_first_bit(&tsk_rt(t)->held_gpus, sizeof(tsk_rt(t)->held_gpus));
849
850 if (!binheap_is_in_heap(&tsk_rt(t)->gpu_owner_node)) {
851 WARN_ON(!is_running(t));
852 TRACE_CUR("nv klmirqd may not inherit from %s/%d on GPU %d\n",
853 t->comm, t->pid, gpu);
854 goto out;
855 }
856
857 TRACE_CUR("task %s/%d on GPU %d decresing priority.\n", t->comm, t->pid, gpu);
858 reg = &NV_DEVICE_REG[gpu];
859
860 hp = container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
861 struct task_struct, rt_param);
862 hp_eff = effective_priority(hp);
863 binheap_delete(&tsk_rt(t)->gpu_owner_node, &reg->owners);
864 binheap_add(&tsk_rt(t)->gpu_owner_node, &reg->owners,
865 struct rt_param, gpu_owner_node);
866
867 if (hp == t) { /* t was originally the hp */
868 struct task_struct *new_hp =
869 container_of(binheap_top_entry(&reg->owners, struct rt_param, gpu_owner_node),
870 struct task_struct, rt_param);
871 if (effective_priority(new_hp) != hp_eff) { /* eff prio. of hp has changed */
872 hp_eff = effective_priority(new_hp);
873 TRACE_CUR("%s/%d is no longer hp on GPU %d.\n", t->comm, t->pid, gpu);
874 retval = gpu_klmirqd_decrease_priority(reg->thread, hp_eff);
875 }
876 }
877
878out:
879 return retval;
880}
881
882
883
884
885
886
887
888
889
890static int __reg_nv_device(int reg_device_id, struct task_struct *t)
891{
892 __set_bit(reg_device_id, &tsk_rt(t)->held_gpus);
893
894 return(0);
895}
896
897static int __clear_reg_nv_device(int de_reg_device_id, struct task_struct *t)
898{
899 __clear_bit(de_reg_device_id, &tsk_rt(t)->held_gpus);
900
901 return(0);
902}
903
904
905int reg_nv_device(int reg_device_id, int reg_action, struct task_struct *t)
906{
907 int ret;
908
909 if((reg_device_id < num_online_gpus()) && (reg_device_id >= 0))
910 {
911 if(reg_action)
912 ret = __reg_nv_device(reg_device_id, t);
913 else
914 ret = __clear_reg_nv_device(reg_device_id, t);
915 }
916 else
917 {
918 ret = -ENODEV;
919 }
920
921 return(ret);
922}
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
947//void pai_check_priority_increase(struct task_struct *t, int reg_device_id)
948//{
949// unsigned long flags;
950// nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id];
951//
952//
953//
954// if(reg->max_prio_owner != t) {
955//
956// raw_spin_lock_irqsave(&reg->lock, flags);
957//
958// if(reg->max_prio_owner != t) {
959// if(litmus->compare(t, reg->max_prio_owner)) {
960// litmus->change_prio_pai_tasklet(reg->max_prio_owner, t);
961// reg->max_prio_owner = t;
962// }
963// }
964//
965// raw_spin_unlock_irqrestore(&reg->lock, flags);
966// }
967//}
968//
969//
970//void pai_check_priority_decrease(struct task_struct *t, int reg_device_id)
971//{
972// unsigned long flags;
973// nv_device_registry_t *reg = &NV_DEVICE_REG[reg_device_id];
974//
975// if(reg->max_prio_owner == t) {
976//
977// raw_spin_lock_irqsave(&reg->lock, flags);
978//
979// if(reg->max_prio_owner == t) {
980// reg->max_prio_owner = find_hp_owner(reg, NULL);
981// if(reg->max_prio_owner != t) {
982// litmus->change_prio_pai_tasklet(t, reg->max_prio_owner);
983// }
984// }
985//
986// raw_spin_unlock_irqrestore(&reg->lock, flags);
987// }
988//}
989#endif
990
991
992
993
994
995//static int __reg_nv_device(int reg_device_id, struct task_struct *t)
996//{
997// int ret = 0;
998// int i;
999// struct task_struct *old_max = NULL;
1000//
1001//
1002// raw_spin_lock_irqsave(&reg->lock, flags);
1003//
1004// if(reg->nr_owners < NV_MAX_SIMULT_USERS) {
1005// TRACE_TASK(t, "registers GPU %d\n", reg_device_id);
1006// for(i = 0; i < NV_MAX_SIMULT_USERS; ++i) {
1007// if(reg->owners[i] == NULL) {
1008// reg->owners[i] = t;
1009//
1010// //if(edf_higher_prio(t, reg->max_prio_owner)) {
1011// if(litmus->compare(t, reg->max_prio_owner)) {
1012// old_max = reg->max_prio_owner;
1013// reg->max_prio_owner = t;
1014//
1015//#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
1016// litmus->change_prio_pai_tasklet(old_max, t);
1017//#endif
1018// }
1019//
1020//#ifdef CONFIG_LITMUS_SOFTIRQD
1021// down_and_set_stat(t, HELD, &tsk_rt(t)->klmirqd_sem);
1022//#endif
1023// ++(reg->nr_owners);
1024//
1025// break;
1026// }
1027// }
1028// }
1029// else
1030// {
1031// TRACE_CUR("%s: device %d is already in use!\n", __FUNCTION__, reg_device_id);
1032// //ret = -EBUSY;
1033// }
1034//
1035// raw_spin_unlock_irqrestore(&reg->lock, flags);
1036//
1037// __set_bit(reg_device_id, &tsk_rt(t)->held_gpus);
1038//
1039// return(ret);
1040//}
1041//
1042//static int __clear_reg_nv_device(int de_reg_device_id, struct task_struct *t)
1043//{
1044// int ret = 0;
1045// int i;
1046// unsigned long flags;
1047// nv_device_registry_t *reg = &NV_DEVICE_REG[de_reg_device_id];
1048//
1049//#ifdef CONFIG_LITMUS_SOFTIRQD
1050// struct task_struct* klmirqd_th = get_klmirqd(de_reg_device_id);
1051//#endif
1052//
1053// if(!test_bit(de_reg_device_id, &tsk_rt(t)->held_gpus)) {
1054// return ret;
1055// }
1056//
1057// raw_spin_lock_irqsave(&reg->lock, flags);
1058//
1059// TRACE_TASK(t, "unregisters GPU %d\n", de_reg_device_id);
1060//
1061// for(i = 0; i < NV_MAX_SIMULT_USERS; ++i) {
1062// if(reg->owners[i] == t) {
1063//#ifdef CONFIG_LITMUS_SOFTIRQD
1064// flush_pending(klmirqd_th, t);
1065//#endif
1066// if(reg->max_prio_owner == t) {
1067// reg->max_prio_owner = find_hp_owner(reg, t);
1068//#ifdef CONFIG_LITMUS_PAI_SOFTIRQD
1069// litmus->change_prio_pai_tasklet(t, reg->max_prio_owner);
1070//#endif
1071// }
1072//
1073//#ifdef CONFIG_LITMUS_SOFTIRQD
1074// up_and_set_stat(t, NOT_HELD, &tsk_rt(t)->klmirqd_sem);
1075//#endif
1076//
1077// reg->owners[i] = NULL;
1078// --(reg->nr_owners);
1079//
1080// break;
1081// }
1082// }
1083//
1084// raw_spin_unlock_irqrestore(&reg->lock, flags);
1085//
1086// __clear_bit(de_reg_device_id, &tsk_rt(t)->held_gpus);
1087//
1088// return(ret);
1089//}
1090//
1091//
1092//int reg_nv_device(int reg_device_id, int reg_action, struct task_struct *t)
1093//{
1094// int ret;
1095//
1096// if((reg_device_id < NV_DEVICE_NUM) && (reg_device_id >= 0))
1097// {
1098// if(reg_action)
1099// ret = __reg_nv_device(reg_device_id, t);
1100// else
1101// ret = __clear_reg_nv_device(reg_device_id, t);
1102// }
1103// else
1104// {
1105// ret = -ENODEV;
1106// }
1107//
1108// return(ret);
1109//}
1110
1111
1112
1113//void lock_nv_registry(u32 target_device_id, unsigned long* flags)
1114//{
1115// BUG_ON(target_device_id >= NV_DEVICE_NUM);
1116//
1117// if(in_interrupt())
1118// TRACE("Locking registry for %d.\n", target_device_id);
1119// else
1120// TRACE_CUR("Locking registry for %d.\n", target_device_id);
1121//
1122// raw_spin_lock_irqsave(&NV_DEVICE_REG[target_device_id].lock, *flags);
1123//}
1124//
1125//void unlock_nv_registry(u32 target_device_id, unsigned long* flags)
1126//{
1127// BUG_ON(target_device_id >= NV_DEVICE_NUM);
1128//
1129// if(in_interrupt())
1130// TRACE("Unlocking registry for %d.\n", target_device_id);
1131// else
1132// TRACE_CUR("Unlocking registry for %d.\n", target_device_id);
1133//
1134// raw_spin_unlock_irqrestore(&NV_DEVICE_REG[target_device_id].lock, *flags);
1135//}
1136
1137
generated by cgit v1.2.2 (git 2.25.0) at 2024-11-09 20:03:59 -0500